i965/miptree: Add a helper for getting an isl_surf from a miptree
[mesa.git] / src / mesa / drivers / dri / i965 / intel_mipmap_tree.c
index 0b432eca7fdd3d00e7fe9c3cea6a0dcc14e597f0..5c439fc9e326d849ad9a666e0e7134a7c24de22c 100644 (file)
@@ -262,7 +262,8 @@ intel_miptree_supports_non_msrt_fast_clear(struct brw_context *brw,
    if (brw->gen >= 9) {
       mesa_format linear_format = _mesa_get_srgb_format_linear(mt->format);
       const uint32_t brw_format = brw_format_for_mesa_format(linear_format);
-      return brw_losslessly_compressible_format(brw, brw_format);
+      return isl_format_supports_lossless_compression(brw->intelScreen->devinfo,
+                                                      brw_format);
    } else
       return true;
 }
@@ -583,7 +584,7 @@ intel_miptree_create_layout(struct brw_context *brw,
    } else if (brw->gen >= 9 && num_samples > 1) {
       layout_flags |= MIPTREE_LAYOUT_FORCE_HALIGN16;
    } else {
-      const bool is_lossless_compressed_aux =
+      const UNUSED bool is_lossless_compressed_aux =
          brw->gen >= 9 && num_samples == 1 &&
          mt->format == MESA_FORMAT_R_UINT32;
 
@@ -798,7 +799,8 @@ intel_miptree_create(struct brw_context *brw,
    /* If this miptree is capable of supporting fast color clears, set
     * fast_clear_state appropriately to ensure that fast clears will occur.
     * Allocation of the MCS miptree will be deferred until the first fast
-    * clear actually occurs.
+    * clear actually occurs or when compressed single sampled buffer is
+    * written by the GPU for the first time.
     */
    if (intel_tiling_supports_non_msrt_mcs(brw, mt->tiling) &&
        intel_miptree_supports_non_msrt_fast_clear(brw, mt)) {
@@ -1022,6 +1024,9 @@ intel_miptree_release(struct intel_mipmap_tree **mt)
       intel_miptree_release(&(*mt)->mcs_mt);
       intel_resolve_map_clear(&(*mt)->hiz_map);
 
+      intel_miptree_release(&(*mt)->plane[0]);
+      intel_miptree_release(&(*mt)->plane[1]);
+
       for (i = 0; i < MAX_TEXTURE_LEVELS; i++) {
         free((*mt)->level[i].slice);
       }
@@ -1603,11 +1608,30 @@ intel_miptree_alloc_non_msrt_mcs(struct brw_context *brw,
    unsigned mcs_height =
       ALIGN(mt->logical_height0, height_divisor) / height_divisor;
    assert(mt->logical_depth0 == 1);
-   uint32_t layout_flags = MIPTREE_LAYOUT_ACCELERATED_UPLOAD |
-                           MIPTREE_LAYOUT_TILING_Y;
+   uint32_t layout_flags = MIPTREE_LAYOUT_TILING_Y;
+
    if (brw->gen >= 8) {
       layout_flags |= MIPTREE_LAYOUT_FORCE_HALIGN16;
    }
+
+   /* On Gen9+ clients are not currently capable of consuming compressed
+    * single-sampled buffers. Disabling compression allows us to skip
+    * resolves.
+    */
+   const bool lossless_compression_disabled = INTEL_DEBUG & DEBUG_NO_RBC;
+   const bool is_lossless_compressed =
+      unlikely(!lossless_compression_disabled) &&
+      brw->gen >= 9 && !mt->is_scanout &&
+      intel_miptree_supports_lossless_compressed(brw, mt);
+
+   /* In case of compression mcs buffer needs to be initialised requiring the
+    * buffer to be immediately mapped to cpu space for writing. Therefore do
+    * not use the gpu access flag which can cause an unnecessary delay if the
+    * backing pages happened to be just used by the GPU.
+    */
+   if (!is_lossless_compressed)
+      layout_flags |= MIPTREE_LAYOUT_ACCELERATED_UPLOAD;
+
    mt->mcs_mt = miptree_create(brw,
                                mt->target,
                                format,
@@ -1619,6 +1643,25 @@ intel_miptree_alloc_non_msrt_mcs(struct brw_context *brw,
                                0 /* num_samples */,
                                layout_flags);
 
+   /* From Gen9 onwards single-sampled (non-msrt) auxiliary buffers are
+    * used for lossless compression which requires similar initialisation
+    * as multi-sample compression.
+    */
+   if (is_lossless_compressed) {
+      /* Hardware sets the auxiliary buffer to all zeroes when it does full
+       * resolve. Initialize it accordingly in case the first renderer is
+       * cpu (or other none compression aware party).
+       *
+       * This is also explicitly stated in the spec (MCS Buffer for Render
+       * Target(s)):
+       *   "If Software wants to enable Color Compression without Fast clear,
+       *    Software needs to initialize MCS with zeros."
+       */
+      intel_miptree_init_mcs(brw, mt, 0);
+      mt->fast_clear_state = INTEL_FAST_CLEAR_STATE_RESOLVED;
+      mt->msaa_layout = INTEL_MSAA_LAYOUT_CMS;
+   }
+
    return mt->mcs_mt;
 }
 
@@ -2236,32 +2279,18 @@ intel_miptree_updownsample(struct brw_context *brw,
                            struct intel_mipmap_tree *src,
                            struct intel_mipmap_tree *dst)
 {
-   /* There is support for only up to eight samples. */
-   const bool use_blorp = src->num_samples <= 8 && dst->num_samples <= 8;
-
-   if (use_blorp) {
-      brw_blorp_blit_miptrees(brw,
-                              src, 0 /* level */, 0 /* layer */,
-                              src->format, SWIZZLE_XYZW,
-                              dst, 0 /* level */, 0 /* layer */, dst->format,
-                              0, 0,
-                              src->logical_width0, src->logical_height0,
-                              0, 0,
-                              dst->logical_width0, dst->logical_height0,
-                              GL_NEAREST, false, false /*mirror x, y*/,
-                              false, false);
-   } else if (src->format == MESA_FORMAT_S_UINT8) {
-      brw_meta_stencil_updownsample(brw, src, dst);
-   } else {
-      brw_meta_updownsample(brw, src, dst);
-   }
+   brw_blorp_blit_miptrees(brw,
+                           src, 0 /* level */, 0 /* layer */,
+                           src->format, SWIZZLE_XYZW,
+                           dst, 0 /* level */, 0 /* layer */, dst->format,
+                           0, 0,
+                           src->logical_width0, src->logical_height0,
+                           0, 0,
+                           dst->logical_width0, dst->logical_height0,
+                           GL_NEAREST, false, false /*mirror x, y*/,
+                           false, false);
 
    if (src->stencil_mt) {
-      if (!use_blorp) {
-         brw_meta_stencil_updownsample(brw, src->stencil_mt, dst);
-         return;
-      }
-
       brw_blorp_blit_miptrees(brw,
                               src->stencil_mt, 0 /* level */, 0 /* layer */,
                               src->stencil_mt->format, SWIZZLE_XYZW,
@@ -2968,3 +2997,173 @@ intel_miptree_unmap(struct brw_context *brw,
 
    intel_miptree_release_map(mt, level, slice);
 }
+
+void
+intel_miptree_get_isl_surf(struct brw_context *brw,
+                           const struct intel_mipmap_tree *mt,
+                           struct isl_surf *surf)
+{
+   switch (mt->target) {
+   case GL_TEXTURE_1D:
+   case GL_TEXTURE_1D_ARRAY: {
+      surf->dim = ISL_SURF_DIM_1D;
+      if (brw->gen >= 9 && mt->tiling == I915_TILING_NONE)
+         surf->dim_layout = ISL_DIM_LAYOUT_GEN9_1D;
+      else
+         surf->dim_layout = ISL_DIM_LAYOUT_GEN4_2D;
+      break;
+   }
+   case GL_TEXTURE_2D:
+   case GL_TEXTURE_2D_ARRAY:
+   case GL_TEXTURE_RECTANGLE:
+   case GL_TEXTURE_CUBE_MAP:
+   case GL_TEXTURE_CUBE_MAP_ARRAY:
+   case GL_TEXTURE_2D_MULTISAMPLE:
+   case GL_TEXTURE_2D_MULTISAMPLE_ARRAY:
+   case GL_TEXTURE_EXTERNAL_OES:
+      surf->dim = ISL_SURF_DIM_2D;
+      surf->dim_layout = ISL_DIM_LAYOUT_GEN4_2D;
+      break;
+   case GL_TEXTURE_3D:
+      surf->dim = ISL_SURF_DIM_3D;
+      if (brw->gen >= 9)
+         surf->dim_layout = ISL_DIM_LAYOUT_GEN4_2D;
+      else
+         surf->dim_layout = ISL_DIM_LAYOUT_GEN4_3D;
+      break;
+   default:
+      unreachable("Invalid texture target");
+   }
+
+   if (mt->num_samples > 1) {
+      switch (mt->msaa_layout) {
+      case INTEL_MSAA_LAYOUT_IMS:
+         surf->msaa_layout = ISL_MSAA_LAYOUT_INTERLEAVED;
+         break;
+      case INTEL_MSAA_LAYOUT_UMS:
+      case INTEL_MSAA_LAYOUT_CMS:
+         surf->msaa_layout = ISL_MSAA_LAYOUT_ARRAY;
+         break;
+      default:
+         unreachable("Invalid MSAA layout");
+      }
+   } else {
+      surf->msaa_layout = ISL_MSAA_LAYOUT_NONE;
+   }
+
+   if (mt->format == MESA_FORMAT_S_UINT8) {
+      surf->tiling = ISL_TILING_W;
+      /* The ISL definition of row_pitch matches the surface state pitch field
+       * a bit better than intel_mipmap_tree.  In particular, ISL incorporates
+       * the factor of 2 for W-tiling in row_pitch.
+       */
+      surf->row_pitch = 2 * mt->pitch;
+   } else {
+      switch (mt->tiling) {
+      case I915_TILING_NONE:
+         surf->tiling = ISL_TILING_LINEAR;
+         break;
+      case I915_TILING_X:
+         surf->tiling = ISL_TILING_X;
+         break;
+      case I915_TILING_Y:
+         switch (mt->tr_mode) {
+         case INTEL_MIPTREE_TRMODE_NONE:
+            surf->tiling = ISL_TILING_Y0;
+            break;
+         case INTEL_MIPTREE_TRMODE_YF:
+            surf->tiling = ISL_TILING_Yf;
+            break;
+         case INTEL_MIPTREE_TRMODE_YS:
+            surf->tiling = ISL_TILING_Ys;
+            break;
+         }
+         break;
+      default:
+         unreachable("Invalid tiling mode");
+      }
+
+      surf->row_pitch = mt->pitch;
+   }
+
+   surf->format = translate_tex_format(brw, mt->format, false);
+
+   if (brw->gen >= 9) {
+      if (surf->dim == ISL_SURF_DIM_1D && surf->tiling == ISL_TILING_LINEAR) {
+         /* For gen9 1-D surfaces, intel_mipmap_tree has a bogus alignment. */
+         surf->image_alignment_el = isl_extent3d(64, 1, 1);
+      } else {
+         /* On gen9+, intel_mipmap_tree stores the horizontal and vertical
+          * alignment in terms of surface elements like we want.
+          */
+         surf->image_alignment_el = isl_extent3d(mt->halign, mt->valign, 1);
+      }
+   } else {
+      /* On earlier gens it's stored in pixels. */
+      unsigned bw, bh;
+      _mesa_get_format_block_size(mt->format, &bw, &bh);
+      surf->image_alignment_el =
+         isl_extent3d(mt->halign / bw, mt->valign / bh, 1);
+   }
+
+   surf->logical_level0_px.width = mt->logical_width0;
+   surf->logical_level0_px.height = mt->logical_height0;
+   if (surf->dim == ISL_SURF_DIM_3D) {
+      surf->logical_level0_px.depth = mt->logical_depth0;
+      surf->logical_level0_px.array_len = 1;
+   } else if (mt->target == GL_TEXTURE_CUBE_MAP ||
+              mt->target == GL_TEXTURE_CUBE_MAP_ARRAY) {
+      /* For cube maps, mt->logical_depth0 is in number of cubes */
+      surf->logical_level0_px.depth = 1;
+      surf->logical_level0_px.array_len = mt->logical_depth0 * 6;
+   } else {
+      surf->logical_level0_px.depth = 1;
+      surf->logical_level0_px.array_len = mt->logical_depth0;
+   }
+
+   surf->phys_level0_sa.width = mt->physical_width0;
+   surf->phys_level0_sa.height = mt->physical_height0;
+   if (surf->dim == ISL_SURF_DIM_3D) {
+      surf->phys_level0_sa.depth = mt->physical_depth0;
+      surf->phys_level0_sa.array_len = 1;
+   } else {
+      surf->phys_level0_sa.depth = 1;
+      surf->phys_level0_sa.array_len = mt->physical_depth0;
+   }
+
+   surf->levels = mt->last_level + 1;
+   surf->samples = MAX2(mt->num_samples, 1);
+
+   surf->size = 0; /* TODO */
+   surf->alignment = 0; /* TODO */
+
+   switch (surf->dim_layout) {
+   case ISL_DIM_LAYOUT_GEN4_2D:
+   case ISL_DIM_LAYOUT_GEN4_3D:
+      if (brw->gen >= 9) {
+         surf->array_pitch_el_rows = mt->qpitch;
+      } else {
+         unsigned bw, bh;
+         _mesa_get_format_block_size(mt->format, &bw, &bh);
+         assert(mt->qpitch % bh == 0);
+         surf->array_pitch_el_rows = mt->qpitch / bh;
+      }
+      break;
+   case ISL_DIM_LAYOUT_GEN9_1D:
+      surf->array_pitch_el_rows = 1;
+      break;
+   }
+
+   switch (mt->array_layout) {
+   case ALL_LOD_IN_EACH_SLICE:
+      surf->array_pitch_span = ISL_ARRAY_PITCH_SPAN_FULL;
+      break;
+   case ALL_SLICES_AT_EACH_LOD:
+      surf->array_pitch_span = ISL_ARRAY_PITCH_SPAN_COMPACT;
+      break;
+   default:
+      unreachable("Invalid array layout");
+   }
+
+   surf->usage = 0; /* TODO */
+}